4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 #include <linux/sched.h>
40 #include <linux/highmem.h>
41 #include <linux/pagemap.h>
43 #define DEBUG_SUBSYSTEM S_LLITE
45 #include <obd_support.h>
46 #include <lustre_lite.h>
47 #include <lustre_dlm.h>
48 #include "llite_internal.h"
50 #define SA_OMITTED_ENTRY_MAX 8ULL
53 /** negative values are for error cases */
54 SA_ENTRY_INIT = 0, /** init entry */
55 SA_ENTRY_SUCC = 1, /** stat succeed */
56 SA_ENTRY_INVA = 2, /** invalid entry */
57 SA_ENTRY_DEST = 3, /** entry to be destroyed */
61 /* link into sai->sai_entries */
62 struct list_head se_link;
63 /* link into sai->sai_entries_{received,stated} */
64 struct list_head se_list;
65 /* link into sai hash table locally */
66 struct list_head se_hash;
67 /* entry reference count */
69 /* entry index in the sai */
71 /* low layer ldlm lock handle */
75 /* entry size, contains name */
77 /* pointer to async getattr enqueue info */
78 struct md_enqueue_info *se_minfo;
79 /* pointer to the async getattr request */
80 struct ptlrpc_request *se_req;
81 /* pointer to the target inode */
82 struct inode *se_inode;
87 static unsigned int sai_generation = 0;
88 static DEFINE_SPINLOCK(sai_generation_lock);
90 static inline int ll_sa_entry_unhashed(struct ll_sa_entry *entry)
92 return list_empty(&entry->se_hash);
96 * The entry only can be released by the caller, it is necessary to hold lock.
98 static inline int ll_sa_entry_stated(struct ll_sa_entry *entry)
101 return (entry->se_stat != SA_ENTRY_INIT);
104 static inline int ll_sa_entry_hash(int val)
106 return val & LL_SA_CACHE_MASK;
110 * Insert entry to hash SA table.
113 ll_sa_entry_enhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
115 int i = ll_sa_entry_hash(entry->se_qstr.hash);
117 spin_lock(&sai->sai_cache_lock[i]);
118 list_add_tail(&entry->se_hash, &sai->sai_cache[i]);
119 spin_unlock(&sai->sai_cache_lock[i]);
123 * Remove entry from SA table.
126 ll_sa_entry_unhash(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
128 int i = ll_sa_entry_hash(entry->se_qstr.hash);
130 spin_lock(&sai->sai_cache_lock[i]);
131 list_del_init(&entry->se_hash);
132 spin_unlock(&sai->sai_cache_lock[i]);
135 static inline int agl_should_run(struct ll_statahead_info *sai,
138 return (inode != NULL && S_ISREG(inode->i_mode) && sai->sai_agl_valid);
141 static inline struct ll_sa_entry *
142 sa_first_received_entry(struct ll_statahead_info *sai)
144 return list_entry(sai->sai_entries_received.next,
145 struct ll_sa_entry, se_list);
148 static inline struct ll_inode_info *
149 agl_first_entry(struct ll_statahead_info *sai)
151 return list_entry(sai->sai_entries_agl.next,
152 struct ll_inode_info, lli_agl_list);
155 static inline int sa_sent_full(struct ll_statahead_info *sai)
157 return atomic_read(&sai->sai_cache_count) >= sai->sai_max;
160 static inline int sa_received_empty(struct ll_statahead_info *sai)
162 return list_empty(&sai->sai_entries_received);
165 static inline int agl_list_empty(struct ll_statahead_info *sai)
167 return list_empty(&sai->sai_entries_agl);
171 * (1) hit ratio less than 80%
173 * (2) consecutive miss more than 8
174 * then means low hit.
176 static inline int sa_low_hit(struct ll_statahead_info *sai)
178 return ((sai->sai_hit > 7 && sai->sai_hit < 4 * sai->sai_miss) ||
179 (sai->sai_consecutive_miss > 8));
183 * If the given index is behind of statahead window more than
184 * SA_OMITTED_ENTRY_MAX, then it is old.
186 static inline int is_omitted_entry(struct ll_statahead_info *sai, __u64 index)
188 return ((__u64)sai->sai_max + index + SA_OMITTED_ENTRY_MAX <
193 * Insert it into sai_entries tail when init.
195 static struct ll_sa_entry *
196 ll_sa_entry_alloc(struct ll_statahead_info *sai, __u64 index,
197 const char *name, int len)
199 struct ll_inode_info *lli;
200 struct ll_sa_entry *entry;
205 entry_size = sizeof(struct ll_sa_entry) + (len & ~3) + 4;
206 OBD_ALLOC(entry, entry_size);
207 if (unlikely(entry == NULL))
208 RETURN(ERR_PTR(-ENOMEM));
210 CDEBUG(D_READA, "alloc sa entry %.*s(%p) index "LPU64"\n",
211 len, name, entry, index);
213 entry->se_index = index;
216 * Statahead entry reference rules:
218 * 1) When statahead entry is initialized, its reference is set as 2.
219 * One reference is used by the directory scanner. When the scanner
220 * searches the statahead cache for the given name, it can perform
221 * lockless hash lookup (only the scanner can remove entry from hash
222 * list), and once found, it needn't to call "atomic_inc()" for the
223 * entry reference. So the performance is improved. After using the
224 * statahead entry, the scanner will call "atomic_dec()" to drop the
225 * reference held when initialization. If it is the last reference,
226 * the statahead entry will be freed.
228 * 2) All other threads, including statahead thread and ptlrpcd thread,
229 * when they process the statahead entry, the reference for target
230 * should be held to guarantee the entry will not be released by the
231 * directory scanner. After processing the entry, these threads will
232 * drop the entry reference. If it is the last reference, the entry
235 * The second reference when initializes the statahead entry is used
236 * by the statahead thread, following the rule 2).
238 atomic_set(&entry->se_refcount, 2);
239 entry->se_stat = SA_ENTRY_INIT;
240 entry->se_size = entry_size;
241 dname = (char *)entry + sizeof(struct ll_sa_entry);
242 memcpy(dname, name, len);
244 entry->se_qstr.hash = full_name_hash(name, len);
245 entry->se_qstr.len = len;
246 entry->se_qstr.name = dname;
248 lli = ll_i2info(sai->sai_inode);
249 spin_lock(&lli->lli_sa_lock);
250 list_add_tail(&entry->se_link, &sai->sai_entries);
251 INIT_LIST_HEAD(&entry->se_list);
252 ll_sa_entry_enhash(sai, entry);
253 spin_unlock(&lli->lli_sa_lock);
255 atomic_inc(&sai->sai_cache_count);
261 * Used by the directory scanner to search entry with name.
263 * Only the caller can remove the entry from hash, so it is unnecessary to hold
264 * hash lock. It is caller's duty to release the init refcount on the entry, so
265 * it is also unnecessary to increase refcount on the entry.
267 static struct ll_sa_entry *
268 ll_sa_entry_get_byname(struct ll_statahead_info *sai, const struct qstr *qstr)
270 struct ll_sa_entry *entry;
271 int i = ll_sa_entry_hash(qstr->hash);
273 list_for_each_entry(entry, &sai->sai_cache[i], se_hash) {
274 if (entry->se_qstr.hash == qstr->hash &&
275 entry->se_qstr.len == qstr->len &&
276 memcmp(entry->se_qstr.name, qstr->name, qstr->len) == 0)
283 * Used by the async getattr request callback to find entry with index.
285 * Inside lli_sa_lock to prevent others to change the list during the search.
286 * It needs to increase entry refcount before returning to guarantee that the
287 * entry cannot be freed by others.
289 static struct ll_sa_entry *
290 ll_sa_entry_get_byindex(struct ll_statahead_info *sai, __u64 index)
292 struct ll_sa_entry *entry;
294 list_for_each_entry(entry, &sai->sai_entries, se_link) {
295 if (entry->se_index == index) {
296 LASSERT(atomic_read(&entry->se_refcount) > 0);
297 atomic_inc(&entry->se_refcount);
300 if (entry->se_index > index)
306 static void ll_sa_entry_cleanup(struct ll_statahead_info *sai,
307 struct ll_sa_entry *entry)
309 struct md_enqueue_info *minfo = entry->se_minfo;
310 struct ptlrpc_request *req = entry->se_req;
313 entry->se_minfo = NULL;
314 ll_intent_release(&minfo->mi_it);
320 entry->se_req = NULL;
321 ptlrpc_req_finished(req);
325 static void ll_sa_entry_put(struct ll_statahead_info *sai,
326 struct ll_sa_entry *entry)
328 if (atomic_dec_and_test(&entry->se_refcount)) {
329 CDEBUG(D_READA, "free sa entry %.*s(%p) index "LPU64"\n",
330 entry->se_qstr.len, entry->se_qstr.name, entry,
333 LASSERT(list_empty(&entry->se_link));
334 LASSERT(list_empty(&entry->se_list));
335 LASSERT(ll_sa_entry_unhashed(entry));
337 ll_sa_entry_cleanup(sai, entry);
339 iput(entry->se_inode);
341 OBD_FREE(entry, entry->se_size);
342 atomic_dec(&sai->sai_cache_count);
347 do_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
349 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
351 LASSERT(!ll_sa_entry_unhashed(entry));
352 LASSERT(!list_empty(&entry->se_link));
354 ll_sa_entry_unhash(sai, entry);
356 spin_lock(&lli->lli_sa_lock);
357 entry->se_stat = SA_ENTRY_DEST;
358 list_del_init(&entry->se_link);
359 if (likely(!list_empty(&entry->se_list)))
360 list_del_init(&entry->se_list);
361 spin_unlock(&lli->lli_sa_lock);
363 ll_sa_entry_put(sai, entry);
367 * Delete it from sai_entries_stated list when fini.
370 ll_sa_entry_fini(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
372 struct ll_sa_entry *pos, *next;
375 do_sa_entry_fini(sai, entry);
377 /* drop old entry, only 'scanner' process does this, no need to lock */
378 list_for_each_entry_safe(pos, next, &sai->sai_entries, se_link) {
379 if (!is_omitted_entry(sai, pos->se_index))
381 do_sa_entry_fini(sai, pos);
386 * Inside lli_sa_lock.
389 do_sa_entry_to_stated(struct ll_statahead_info *sai,
390 struct ll_sa_entry *entry, se_stat_t stat)
392 struct ll_sa_entry *se;
393 struct list_head *pos = &sai->sai_entries_stated;
395 if (!list_empty(&entry->se_list))
396 list_del_init(&entry->se_list);
398 list_for_each_entry_reverse(se, &sai->sai_entries_stated, se_list) {
399 if (se->se_index < entry->se_index) {
405 list_add(&entry->se_list, pos);
406 entry->se_stat = stat;
410 * Move entry to sai_entries_stated and sort with the index.
411 * \retval 1 -- entry to be destroyed.
412 * \retval 0 -- entry is inserted into stated list.
415 ll_sa_entry_to_stated(struct ll_statahead_info *sai,
416 struct ll_sa_entry *entry, se_stat_t stat)
418 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
421 ll_sa_entry_cleanup(sai, entry);
423 spin_lock(&lli->lli_sa_lock);
424 if (likely(entry->se_stat != SA_ENTRY_DEST)) {
425 do_sa_entry_to_stated(sai, entry, stat);
428 spin_unlock(&lli->lli_sa_lock);
434 * Insert inode into the list of sai_entries_agl.
436 static void ll_agl_add(struct ll_statahead_info *sai,
437 struct inode *inode, int index)
439 struct ll_inode_info *child = ll_i2info(inode);
440 struct ll_inode_info *parent = ll_i2info(sai->sai_inode);
443 spin_lock(&child->lli_agl_lock);
444 if (child->lli_agl_index == 0) {
445 child->lli_agl_index = index;
446 spin_unlock(&child->lli_agl_lock);
448 LASSERT(list_empty(&child->lli_agl_list));
451 spin_lock(&parent->lli_agl_lock);
452 if (agl_list_empty(sai))
454 list_add_tail(&child->lli_agl_list, &sai->sai_entries_agl);
455 spin_unlock(&parent->lli_agl_lock);
457 spin_unlock(&child->lli_agl_lock);
461 wake_up(&sai->sai_agl_thread.t_ctl_waitq);
464 static struct ll_statahead_info *ll_sai_alloc(void)
466 struct ll_statahead_info *sai;
474 atomic_set(&sai->sai_refcount, 1);
476 spin_lock(&sai_generation_lock);
477 sai->sai_generation = ++sai_generation;
478 if (unlikely(sai_generation == 0))
479 sai->sai_generation = ++sai_generation;
480 spin_unlock(&sai_generation_lock);
482 sai->sai_max = LL_SA_RPC_MIN;
484 init_waitqueue_head(&sai->sai_waitq);
485 init_waitqueue_head(&sai->sai_thread.t_ctl_waitq);
486 init_waitqueue_head(&sai->sai_agl_thread.t_ctl_waitq);
488 INIT_LIST_HEAD(&sai->sai_entries);
489 INIT_LIST_HEAD(&sai->sai_entries_received);
490 INIT_LIST_HEAD(&sai->sai_entries_stated);
491 INIT_LIST_HEAD(&sai->sai_entries_agl);
493 for (i = 0; i < LL_SA_CACHE_SIZE; i++) {
494 INIT_LIST_HEAD(&sai->sai_cache[i]);
495 spin_lock_init(&sai->sai_cache_lock[i]);
497 atomic_set(&sai->sai_cache_count, 0);
502 static inline struct ll_statahead_info *
503 ll_sai_get(struct ll_statahead_info *sai)
505 atomic_inc(&sai->sai_refcount);
509 static void ll_sai_put(struct ll_statahead_info *sai)
511 struct inode *inode = sai->sai_inode;
512 struct ll_inode_info *lli = ll_i2info(inode);
515 if (atomic_dec_and_lock(&sai->sai_refcount, &lli->lli_sa_lock)) {
516 struct ll_sa_entry *entry, *next;
518 if (unlikely(atomic_read(&sai->sai_refcount) > 0)) {
519 /* It is race case, the interpret callback just hold
520 * a reference count */
521 spin_unlock(&lli->lli_sa_lock);
525 LASSERT(lli->lli_opendir_key == NULL);
526 LASSERT(thread_is_stopped(&sai->sai_thread));
527 LASSERT(thread_is_stopped(&sai->sai_agl_thread));
530 lli->lli_opendir_pid = 0;
531 spin_unlock(&lli->lli_sa_lock);
533 if (sai->sai_sent > sai->sai_replied)
534 CDEBUG(D_READA,"statahead for dir "DFID" does not "
535 "finish: [sent:"LPU64"] [replied:"LPU64"]\n",
537 sai->sai_sent, sai->sai_replied);
539 list_for_each_entry_safe(entry, next,
540 &sai->sai_entries, se_link)
541 do_sa_entry_fini(sai, entry);
543 LASSERT(list_empty(&sai->sai_entries));
544 LASSERT(sa_received_empty(sai));
545 LASSERT(list_empty(&sai->sai_entries_stated));
547 LASSERT(atomic_read(&sai->sai_cache_count) == 0);
548 LASSERT(agl_list_empty(sai));
557 /* Do NOT forget to drop inode refcount when into sai_entries_agl. */
558 static void ll_agl_trigger(struct inode *inode, struct ll_statahead_info *sai)
560 struct ll_inode_info *lli = ll_i2info(inode);
561 __u64 index = lli->lli_agl_index;
565 LASSERT(list_empty(&lli->lli_agl_list));
567 /* AGL maybe fall behind statahead with one entry */
568 if (is_omitted_entry(sai, index + 1)) {
569 lli->lli_agl_index = 0;
574 /* Someone is in glimpse (sync or async), do nothing. */
575 rc = down_write_trylock(&lli->lli_glimpse_sem);
577 lli->lli_agl_index = 0;
583 * Someone triggered glimpse within 1 sec before.
584 * 1) The former glimpse succeeded with glimpse lock granted by OST, and
585 * if the lock is still cached on client, AGL needs to do nothing. If
586 * it is cancelled by other client, AGL maybe cannot obtaion new lock
587 * for no glimpse callback triggered by AGL.
588 * 2) The former glimpse succeeded, but OST did not grant glimpse lock.
589 * Under such case, it is quite possible that the OST will not grant
590 * glimpse lock for AGL also.
591 * 3) The former glimpse failed, compared with other two cases, it is
592 * relative rare. AGL can ignore such case, and it will not muchly
593 * affect the performance.
595 if (lli->lli_glimpse_time != 0 &&
596 cfs_time_before(cfs_time_shift(-1), lli->lli_glimpse_time)) {
597 up_write(&lli->lli_glimpse_sem);
598 lli->lli_agl_index = 0;
603 CDEBUG(D_READA, "Handling (init) async glimpse: inode = "
604 DFID", idx = "LPU64"\n", PFID(&lli->lli_fid), index);
607 lli->lli_agl_index = 0;
608 lli->lli_glimpse_time = cfs_time_current();
609 up_write(&lli->lli_glimpse_sem);
611 CDEBUG(D_READA, "Handled (init) async glimpse: inode= "
612 DFID", idx = "LPU64", rc = %d\n",
613 PFID(&lli->lli_fid), index, rc);
620 static void ll_post_statahead(struct ll_statahead_info *sai)
622 struct inode *dir = sai->sai_inode;
624 struct ll_inode_info *lli = ll_i2info(dir);
625 struct ll_sa_entry *entry;
626 struct md_enqueue_info *minfo;
627 struct lookup_intent *it;
628 struct ptlrpc_request *req;
629 struct mdt_body *body;
633 spin_lock(&lli->lli_sa_lock);
634 if (unlikely(sa_received_empty(sai))) {
635 spin_unlock(&lli->lli_sa_lock);
638 entry = sa_first_received_entry(sai);
639 atomic_inc(&entry->se_refcount);
640 list_del_init(&entry->se_list);
641 spin_unlock(&lli->lli_sa_lock);
643 LASSERT(entry->se_handle != 0);
645 minfo = entry->se_minfo;
648 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
650 GOTO(out, rc = -EFAULT);
652 child = entry->se_inode;
657 LASSERT(fid_is_zero(&minfo->mi_data.op_fid2));
659 /* XXX: No fid in reply, this is probaly cross-ref case.
660 * SA can't handle it yet. */
661 if (body->mbo_valid & OBD_MD_MDS)
662 GOTO(out, rc = -EAGAIN);
667 /* unlinked and re-created with the same name */
668 if (unlikely(!lu_fid_eq(&minfo->mi_data.op_fid2,
670 entry->se_inode = NULL;
676 it->d.lustre.it_lock_handle = entry->se_handle;
677 rc = md_revalidate_lock(ll_i2mdexp(dir), it, ll_inode2fid(dir), NULL);
679 GOTO(out, rc = -EAGAIN);
681 rc = ll_prep_inode(&child, req, dir->i_sb, it);
685 CDEBUG(D_DLMTRACE, "%s: setting l_data to inode "DFID"(%p)\n",
686 ll_get_fsname(child->i_sb, NULL, 0),
687 PFID(ll_inode2fid(child)), child);
688 ll_set_lock_data(ll_i2sbi(dir)->ll_md_exp, child, it, NULL);
690 entry->se_inode = child;
692 if (agl_should_run(sai, child))
693 ll_agl_add(sai, child, entry->se_index);
698 /* The "ll_sa_entry_to_stated()" will drop related ldlm ibits lock
699 * reference count by calling "ll_intent_drop_lock()" in spite of the
700 * above operations failed or not. Do not worry about calling
701 * "ll_intent_drop_lock()" more than once. */
702 rc = ll_sa_entry_to_stated(sai, entry,
703 rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
704 if (rc == 0 && entry->se_index == sai->sai_index_wait)
705 wake_up(&sai->sai_waitq);
706 ll_sa_entry_put(sai, entry);
709 static int ll_statahead_interpret(struct ptlrpc_request *req,
710 struct md_enqueue_info *minfo, int rc)
712 struct lookup_intent *it = &minfo->mi_it;
713 struct inode *dir = minfo->mi_dir;
714 struct ll_inode_info *lli = ll_i2info(dir);
715 struct ll_statahead_info *sai = NULL;
716 struct ll_sa_entry *entry;
720 if (it_disposition(it, DISP_LOOKUP_NEG))
723 spin_lock(&lli->lli_sa_lock);
725 if (unlikely(lli->lli_sai == NULL ||
726 lli->lli_sai->sai_generation != minfo->mi_generation)) {
727 spin_unlock(&lli->lli_sa_lock);
728 GOTO(out, rc = -ESTALE);
730 sai = ll_sai_get(lli->lli_sai);
731 if (unlikely(!thread_is_running(&sai->sai_thread))) {
733 spin_unlock(&lli->lli_sa_lock);
734 GOTO(out, rc = -EBADFD);
737 entry = ll_sa_entry_get_byindex(sai, minfo->mi_cbdata);
740 spin_unlock(&lli->lli_sa_lock);
741 GOTO(out, rc = -EIDRM);
745 do_sa_entry_to_stated(sai, entry, SA_ENTRY_INVA);
746 wakeup = (entry->se_index == sai->sai_index_wait);
748 entry->se_minfo = minfo;
749 entry->se_req = ptlrpc_request_addref(req);
750 /* Release the async ibits lock ASAP to avoid deadlock
751 * when statahead thread tries to enqueue lock on parent
752 * for readpage and other tries to enqueue lock on child
753 * with parent's lock held, for example: unlink. */
754 entry->se_handle = it->d.lustre.it_lock_handle;
755 ll_intent_drop_lock(it);
756 wakeup = sa_received_empty(sai);
757 list_add_tail(&entry->se_list,
758 &sai->sai_entries_received);
761 spin_unlock(&lli->lli_sa_lock);
763 ll_sa_entry_put(sai, entry);
765 wake_up(&sai->sai_thread.t_ctl_waitq);
772 ll_intent_release(it);
781 static void sa_args_fini(struct md_enqueue_info *minfo,
782 struct ldlm_enqueue_info *einfo)
784 LASSERT(minfo && einfo);
786 capa_put(minfo->mi_data.op_capa1);
787 capa_put(minfo->mi_data.op_capa2);
793 * There is race condition between "capa_put" and "ll_statahead_interpret" for
794 * accessing "op_data.op_capa[1,2]" as following:
795 * "capa_put" releases "op_data.op_capa[1,2]"'s reference count after calling
796 * "md_intent_getattr_async". But "ll_statahead_interpret" maybe run first, and
797 * fill "op_data.op_capa[1,2]" as POISON, then cause "capa_put" access invalid
798 * "ocapa". So here reserve "op_data.op_capa[1,2]" in "pcapa" before calling
799 * "md_intent_getattr_async".
801 static int sa_args_init(struct inode *dir, struct inode *child,
802 struct ll_sa_entry *entry, struct md_enqueue_info **pmi,
803 struct ldlm_enqueue_info **pei,
804 struct obd_capa **pcapa)
806 struct qstr *qstr = &entry->se_qstr;
807 struct ll_inode_info *lli = ll_i2info(dir);
808 struct md_enqueue_info *minfo;
809 struct ldlm_enqueue_info *einfo;
810 struct md_op_data *op_data;
812 OBD_ALLOC_PTR(einfo);
816 OBD_ALLOC_PTR(minfo);
822 op_data = ll_prep_md_op_data(&minfo->mi_data, dir, child, qstr->name,
823 qstr->len, 0, LUSTRE_OPC_ANY, NULL);
824 if (IS_ERR(op_data)) {
827 return PTR_ERR(op_data);
830 minfo->mi_it.it_op = IT_GETATTR;
831 minfo->mi_dir = igrab(dir);
832 minfo->mi_cb = ll_statahead_interpret;
833 minfo->mi_generation = lli->lli_sai->sai_generation;
834 minfo->mi_cbdata = entry->se_index;
836 einfo->ei_type = LDLM_IBITS;
837 einfo->ei_mode = it_to_lock_mode(&minfo->mi_it);
838 einfo->ei_cb_bl = ll_md_blocking_ast;
839 einfo->ei_cb_cp = ldlm_completion_ast;
840 einfo->ei_cb_gl = NULL;
841 einfo->ei_cbdata = NULL;
845 pcapa[0] = op_data->op_capa1;
846 pcapa[1] = op_data->op_capa2;
851 static int do_sa_lookup(struct inode *dir, struct ll_sa_entry *entry)
853 struct md_enqueue_info *minfo;
854 struct ldlm_enqueue_info *einfo;
855 struct obd_capa *capas[2];
859 rc = sa_args_init(dir, NULL, entry, &minfo, &einfo, capas);
863 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
868 sa_args_fini(minfo, einfo);
875 * similar to ll_revalidate_it().
876 * \retval 1 -- dentry valid
877 * \retval 0 -- will send stat-ahead request
878 * \retval others -- prepare stat-ahead request failed
880 static int do_sa_revalidate(struct inode *dir, struct ll_sa_entry *entry,
881 struct dentry *dentry)
883 struct inode *inode = dentry->d_inode;
884 struct lookup_intent it = { .it_op = IT_GETATTR,
885 .d.lustre.it_lock_handle = 0 };
886 struct md_enqueue_info *minfo;
887 struct ldlm_enqueue_info *einfo;
888 struct obd_capa *capas[2];
892 if (unlikely(inode == NULL))
895 if (d_mountpoint(dentry))
898 entry->se_inode = igrab(inode);
899 rc = md_revalidate_lock(ll_i2mdexp(dir), &it, ll_inode2fid(inode),NULL);
901 entry->se_handle = it.d.lustre.it_lock_handle;
902 ll_intent_release(&it);
906 rc = sa_args_init(dir, inode, entry, &minfo, &einfo, capas);
908 entry->se_inode = NULL;
913 rc = md_intent_getattr_async(ll_i2mdexp(dir), minfo, einfo);
918 entry->se_inode = NULL;
920 sa_args_fini(minfo, einfo);
926 static void ll_statahead_one(struct dentry *parent, const char* entry_name,
929 struct inode *dir = parent->d_inode;
930 struct ll_inode_info *lli = ll_i2info(dir);
931 struct ll_statahead_info *sai = lli->lli_sai;
932 struct dentry *dentry = NULL;
933 struct ll_sa_entry *entry;
938 entry = ll_sa_entry_alloc(sai, sai->sai_index, entry_name,
943 dentry = d_lookup(parent, &entry->se_qstr);
945 rc = do_sa_lookup(dir, entry);
947 rc = do_sa_revalidate(dir, entry, dentry);
948 if (rc == 1 && agl_should_run(sai, dentry->d_inode))
949 ll_agl_add(sai, dentry->d_inode, entry->se_index);
956 rc1 = ll_sa_entry_to_stated(sai, entry,
957 rc < 0 ? SA_ENTRY_INVA : SA_ENTRY_SUCC);
958 if (rc1 == 0 && entry->se_index == sai->sai_index_wait)
959 wake_up(&sai->sai_waitq);
965 /* drop one refcount on entry by ll_sa_entry_alloc */
966 ll_sa_entry_put(sai, entry);
971 static int ll_agl_thread(void *arg)
973 struct dentry *parent = (struct dentry *)arg;
974 struct inode *dir = parent->d_inode;
975 struct ll_inode_info *plli = ll_i2info(dir);
976 struct ll_inode_info *clli;
977 struct ll_sb_info *sbi = ll_i2sbi(dir);
978 struct ll_statahead_info *sai = ll_sai_get(plli->lli_sai);
979 struct ptlrpc_thread *thread = &sai->sai_agl_thread;
980 struct l_wait_info lwi = { 0 };
983 thread->t_pid = current_pid();
984 CDEBUG(D_READA, "agl thread started: sai %p, parent %.*s\n",
985 sai, parent->d_name.len, parent->d_name.name);
987 atomic_inc(&sbi->ll_agl_total);
988 spin_lock(&plli->lli_agl_lock);
989 sai->sai_agl_valid = 1;
990 if (thread_is_init(thread))
991 /* If someone else has changed the thread state
992 * (e.g. already changed to SVC_STOPPING), we can't just
993 * blindly overwrite that setting. */
994 thread_set_flags(thread, SVC_RUNNING);
995 spin_unlock(&plli->lli_agl_lock);
996 wake_up(&thread->t_ctl_waitq);
999 l_wait_event(thread->t_ctl_waitq,
1000 !agl_list_empty(sai) ||
1001 !thread_is_running(thread),
1004 if (!thread_is_running(thread))
1007 spin_lock(&plli->lli_agl_lock);
1008 /* The statahead thread maybe help to process AGL entries,
1009 * so check whether list empty again. */
1010 if (!agl_list_empty(sai)) {
1011 clli = agl_first_entry(sai);
1012 list_del_init(&clli->lli_agl_list);
1013 spin_unlock(&plli->lli_agl_lock);
1014 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1016 spin_unlock(&plli->lli_agl_lock);
1020 spin_lock(&plli->lli_agl_lock);
1021 sai->sai_agl_valid = 0;
1022 while (!agl_list_empty(sai)) {
1023 clli = agl_first_entry(sai);
1024 list_del_init(&clli->lli_agl_list);
1025 spin_unlock(&plli->lli_agl_lock);
1026 clli->lli_agl_index = 0;
1027 iput(&clli->lli_vfs_inode);
1028 spin_lock(&plli->lli_agl_lock);
1030 thread_set_flags(thread, SVC_STOPPED);
1031 spin_unlock(&plli->lli_agl_lock);
1032 wake_up(&thread->t_ctl_waitq);
1034 CDEBUG(D_READA, "agl thread stopped: sai %p, parent %.*s\n",
1035 sai, parent->d_name.len, parent->d_name.name);
1039 static void ll_start_agl(struct dentry *parent, struct ll_statahead_info *sai)
1041 struct ptlrpc_thread *thread = &sai->sai_agl_thread;
1042 struct l_wait_info lwi = { 0 };
1043 struct ll_inode_info *plli;
1044 struct task_struct *task;
1047 CDEBUG(D_READA, "start agl thread: sai %p, parent %.*s\n",
1048 sai, parent->d_name.len, parent->d_name.name);
1050 plli = ll_i2info(parent->d_inode);
1051 task = kthread_run(ll_agl_thread, parent,
1052 "ll_agl_%u", plli->lli_opendir_pid);
1054 CERROR("can't start ll_agl thread, rc: %ld\n", PTR_ERR(task));
1055 thread_set_flags(thread, SVC_STOPPED);
1059 l_wait_event(thread->t_ctl_waitq,
1060 thread_is_running(thread) || thread_is_stopped(thread),
1065 static int ll_statahead_thread(void *arg)
1067 struct dentry *parent = (struct dentry *)arg;
1068 struct inode *dir = parent->d_inode;
1069 struct ll_inode_info *plli = ll_i2info(dir);
1070 struct ll_inode_info *clli;
1071 struct ll_sb_info *sbi = ll_i2sbi(dir);
1072 struct ll_statahead_info *sai = ll_sai_get(plli->lli_sai);
1073 struct ptlrpc_thread *thread = &sai->sai_thread;
1074 struct ptlrpc_thread *agl_thread = &sai->sai_agl_thread;
1077 struct md_op_data *op_data;
1078 struct ll_dir_chain chain;
1079 struct l_wait_info lwi = { 0 };
1080 struct page *page = NULL;
1084 thread->t_pid = current_pid();
1085 CDEBUG(D_READA, "statahead thread starting: sai %p, parent %.*s\n",
1086 sai, parent->d_name.len, parent->d_name.name);
1088 op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1089 LUSTRE_OPC_ANY, dir);
1090 if (IS_ERR(op_data))
1091 GOTO(out_put, rc = PTR_ERR(op_data));
1093 op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1095 if (sbi->ll_flags & LL_SBI_AGL_ENABLED)
1096 ll_start_agl(parent, sai);
1098 atomic_inc(&sbi->ll_sa_total);
1099 spin_lock(&plli->lli_sa_lock);
1100 if (thread_is_init(thread))
1101 /* If someone else has changed the thread state
1102 * (e.g. already changed to SVC_STOPPING), we can't just
1103 * blindly overwrite that setting. */
1104 thread_set_flags(thread, SVC_RUNNING);
1105 spin_unlock(&plli->lli_sa_lock);
1106 wake_up(&thread->t_ctl_waitq);
1108 ll_dir_chain_init(&chain);
1109 page = ll_get_dir_page(dir, op_data, pos, &chain);
1111 struct lu_dirpage *dp;
1112 struct lu_dirent *ent;
1116 CDEBUG(D_READA, "error reading dir "DFID" at "LPU64
1117 "/"LPU64" opendir_pid = %u: rc = %d\n",
1118 PFID(ll_inode2fid(dir)), pos, sai->sai_index,
1119 plli->lli_opendir_pid, rc);
1123 dp = page_address(page);
1124 for (ent = lu_dirent_start(dp); ent != NULL;
1125 ent = lu_dirent_next(ent)) {
1130 hash = le64_to_cpu(ent->lde_hash);
1131 if (unlikely(hash < pos))
1133 * Skip until we find target hash value.
1137 namelen = le16_to_cpu(ent->lde_namelen);
1138 if (unlikely(namelen == 0))
1140 * Skip dummy record.
1144 name = ent->lde_name;
1145 if (name[0] == '.') {
1151 } else if (name[1] == '.' && namelen == 2) {
1156 } else if (!sai->sai_ls_all) {
1158 * skip hidden files.
1160 sai->sai_skip_hidden++;
1166 * don't stat-ahead first entry.
1168 if (unlikely(++first == 1))
1172 l_wait_event(thread->t_ctl_waitq,
1173 !sa_sent_full(sai) ||
1174 !sa_received_empty(sai) ||
1175 !agl_list_empty(sai) ||
1176 !thread_is_running(thread),
1180 while (!sa_received_empty(sai))
1181 ll_post_statahead(sai);
1183 if (unlikely(!thread_is_running(thread))) {
1184 ll_release_page(dir, page, false);
1188 /* If no window for metadata statahead, but there are
1189 * some AGL entries to be triggered, then try to help
1190 * to process the AGL entries. */
1191 if (sa_sent_full(sai)) {
1192 spin_lock(&plli->lli_agl_lock);
1193 while (!agl_list_empty(sai)) {
1194 clli = agl_first_entry(sai);
1195 list_del_init(&clli->lli_agl_list);
1196 spin_unlock(&plli->lli_agl_lock);
1197 ll_agl_trigger(&clli->lli_vfs_inode,
1200 if (!sa_received_empty(sai))
1204 !thread_is_running(thread))) {
1205 ll_release_page(dir, page,
1210 if (!sa_sent_full(sai))
1213 spin_lock(&plli->lli_agl_lock);
1215 spin_unlock(&plli->lli_agl_lock);
1220 ll_statahead_one(parent, name, namelen);
1223 pos = le64_to_cpu(dp->ldp_hash_end);
1224 if (pos == MDS_DIR_END_OFF) {
1226 * End of directory reached.
1228 ll_release_page(dir, page, false);
1230 l_wait_event(thread->t_ctl_waitq,
1231 !sa_received_empty(sai) ||
1232 sai->sai_sent == sai->sai_replied ||
1233 !thread_is_running(thread),
1236 while (!sa_received_empty(sai))
1237 ll_post_statahead(sai);
1239 if (unlikely(!thread_is_running(thread)))
1242 if (sai->sai_sent == sai->sai_replied &&
1243 sa_received_empty(sai))
1247 spin_lock(&plli->lli_agl_lock);
1248 while (!agl_list_empty(sai) &&
1249 thread_is_running(thread)) {
1250 clli = agl_first_entry(sai);
1251 list_del_init(&clli->lli_agl_list);
1252 spin_unlock(&plli->lli_agl_lock);
1253 ll_agl_trigger(&clli->lli_vfs_inode, sai);
1254 spin_lock(&plli->lli_agl_lock);
1256 spin_unlock(&plli->lli_agl_lock);
1261 * chain is exhausted.
1262 * Normal case: continue to the next page.
1264 ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1266 sai->sai_in_readpage = 1;
1267 page = ll_get_dir_page(dir, op_data, pos, &chain);
1268 sai->sai_in_readpage = 0;
1273 ll_dir_chain_fini(&chain);
1274 ll_finish_md_op_data(op_data);
1276 if (sai->sai_agl_valid) {
1277 spin_lock(&plli->lli_agl_lock);
1278 thread_set_flags(agl_thread, SVC_STOPPING);
1279 spin_unlock(&plli->lli_agl_lock);
1280 wake_up(&agl_thread->t_ctl_waitq);
1282 CDEBUG(D_READA, "stop agl thread: sai %p pid %u\n",
1283 sai, (unsigned int)agl_thread->t_pid);
1284 l_wait_event(agl_thread->t_ctl_waitq,
1285 thread_is_stopped(agl_thread),
1288 /* Set agl_thread flags anyway. */
1289 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1291 spin_lock(&plli->lli_sa_lock);
1292 if (!sa_received_empty(sai)) {
1293 thread_set_flags(thread, SVC_STOPPING);
1294 spin_unlock(&plli->lli_sa_lock);
1296 /* To release the resources held by received entries. */
1297 while (!sa_received_empty(sai))
1298 ll_post_statahead(sai);
1300 spin_lock(&plli->lli_sa_lock);
1302 thread_set_flags(thread, SVC_STOPPED);
1303 spin_unlock(&plli->lli_sa_lock);
1304 wake_up(&sai->sai_waitq);
1305 wake_up(&thread->t_ctl_waitq);
1307 CDEBUG(D_READA, "statahead thread stopped: sai %p, parent %.*s\n",
1308 sai, parent->d_name.len, parent->d_name.name);
1314 * called in ll_file_release().
1316 void ll_stop_statahead(struct inode *dir, void *key)
1318 struct ll_inode_info *lli = ll_i2info(dir);
1320 if (unlikely(key == NULL))
1323 spin_lock(&lli->lli_sa_lock);
1324 if (lli->lli_opendir_key != key || lli->lli_opendir_pid == 0) {
1325 spin_unlock(&lli->lli_sa_lock);
1329 lli->lli_opendir_key = NULL;
1332 struct l_wait_info lwi = { 0 };
1333 struct ptlrpc_thread *thread = &lli->lli_sai->sai_thread;
1335 if (!thread_is_stopped(thread)) {
1336 thread_set_flags(thread, SVC_STOPPING);
1337 spin_unlock(&lli->lli_sa_lock);
1338 wake_up(&thread->t_ctl_waitq);
1340 CDEBUG(D_READA, "stop statahead thread: sai %p pid %u\n",
1341 lli->lli_sai, (unsigned int)thread->t_pid);
1342 l_wait_event(thread->t_ctl_waitq,
1343 thread_is_stopped(thread),
1346 spin_unlock(&lli->lli_sa_lock);
1350 * Put the ref which was held when first statahead_enter.
1351 * It maybe not the last ref for some statahead requests
1354 ll_sai_put(lli->lli_sai);
1356 lli->lli_opendir_pid = 0;
1357 spin_unlock(&lli->lli_sa_lock);
1363 * not first dirent, or is "."
1365 LS_NONE_FIRST_DE = 0,
1367 * the first non-hidden dirent
1371 * the first hidden dirent, that is "."
1376 static int is_first_dirent(struct inode *dir, struct dentry *dentry)
1378 struct ll_dir_chain chain;
1379 struct qstr *target = &dentry->d_name;
1380 struct md_op_data *op_data;
1382 struct page *page = NULL;
1383 int rc = LS_NONE_FIRST_DE;
1387 op_data = ll_prep_md_op_data(NULL, dir, dir, NULL, 0, 0,
1388 LUSTRE_OPC_ANY, dir);
1389 if (IS_ERR(op_data))
1390 RETURN(PTR_ERR(op_data));
1392 *FIXME choose the start offset of the readdir
1394 op_data->op_stripe_offset = 0;
1395 op_data->op_max_pages = ll_i2sbi(dir)->ll_md_brw_pages;
1397 ll_dir_chain_init(&chain);
1398 page = ll_get_dir_page(dir, op_data, 0, &chain);
1401 struct lu_dirpage *dp;
1402 struct lu_dirent *ent;
1405 struct ll_inode_info *lli = ll_i2info(dir);
1408 CERROR("%s: reading dir "DFID" at "LPU64
1409 "opendir_pid = %u : rc = %d\n",
1410 ll_get_fsname(dir->i_sb, NULL, 0),
1411 PFID(ll_inode2fid(dir)), pos,
1412 lli->lli_opendir_pid, rc);
1416 dp = page_address(page);
1417 for (ent = lu_dirent_start(dp); ent != NULL;
1418 ent = lu_dirent_next(ent)) {
1423 hash = le64_to_cpu(ent->lde_hash);
1424 /* The ll_get_dir_page() can return any page containing
1425 * the given hash which may be not the start hash. */
1426 if (unlikely(hash < pos))
1429 namelen = le16_to_cpu(ent->lde_namelen);
1430 if (unlikely(namelen == 0))
1432 * skip dummy record.
1436 name = ent->lde_name;
1437 if (name[0] == '.') {
1443 else if (name[1] == '.' && namelen == 2)
1454 if (dot_de && target->name[0] != '.') {
1455 CDEBUG(D_READA, "%.*s skip hidden file %.*s\n",
1456 target->len, target->name,
1461 if (target->len != namelen ||
1462 memcmp(target->name, name, namelen) != 0)
1463 rc = LS_NONE_FIRST_DE;
1467 rc = LS_FIRST_DOT_DE;
1469 ll_release_page(dir, page, false);
1472 pos = le64_to_cpu(dp->ldp_hash_end);
1473 if (pos == MDS_DIR_END_OFF) {
1475 * End of directory reached.
1477 ll_release_page(dir, page, false);
1481 * chain is exhausted
1482 * Normal case: continue to the next page.
1484 ll_release_page(dir, page, le32_to_cpu(dp->ldp_flags) &
1486 page = ll_get_dir_page(dir, op_data, pos, &chain);
1491 ll_dir_chain_fini(&chain);
1492 ll_finish_md_op_data(op_data);
1497 ll_sai_unplug(struct ll_statahead_info *sai, struct ll_sa_entry *entry)
1499 struct ptlrpc_thread *thread = &sai->sai_thread;
1500 struct ll_sb_info *sbi = ll_i2sbi(sai->sai_inode);
1504 if (entry != NULL && entry->se_stat == SA_ENTRY_SUCC)
1509 ll_sa_entry_fini(sai, entry);
1512 sai->sai_consecutive_miss = 0;
1513 sai->sai_max = min(2 * sai->sai_max, sbi->ll_sa_max);
1515 struct ll_inode_info *lli = ll_i2info(sai->sai_inode);
1518 sai->sai_consecutive_miss++;
1519 if (sa_low_hit(sai) && thread_is_running(thread)) {
1520 atomic_inc(&sbi->ll_sa_wrong);
1521 CDEBUG(D_READA, "Statahead for dir "DFID" hit "
1522 "ratio too low: hit/miss "LPU64"/"LPU64
1523 ", sent/replied "LPU64"/"LPU64", stopping "
1524 "statahead thread\n",
1525 PFID(&lli->lli_fid), sai->sai_hit,
1526 sai->sai_miss, sai->sai_sent,
1528 spin_lock(&lli->lli_sa_lock);
1529 if (!thread_is_stopped(thread))
1530 thread_set_flags(thread, SVC_STOPPING);
1531 spin_unlock(&lli->lli_sa_lock);
1535 if (!thread_is_stopped(thread))
1536 wake_up(&thread->t_ctl_waitq);
1542 * Start statahead thread if this is the first dir entry.
1543 * Otherwise if a thread is started already, wait it until it is ahead of me.
1544 * \retval 1 -- find entry with lock in cache, the caller needs to do
1546 * \retval 0 -- find entry in cache, but without lock, the caller needs
1548 * \retval others -- the caller need to process as non-statahead.
1550 int do_statahead_enter(struct inode *dir, struct dentry **dentryp,
1553 struct ll_inode_info *lli = ll_i2info(dir);
1554 struct ll_statahead_info *sai = lli->lli_sai;
1555 struct dentry *parent;
1556 struct ll_sa_entry *entry;
1557 struct ptlrpc_thread *thread;
1558 struct l_wait_info lwi = { 0 };
1559 struct task_struct *task;
1561 struct ll_inode_info *plli;
1564 LASSERT(lli->lli_opendir_pid == current_pid());
1567 thread = &sai->sai_thread;
1568 if (unlikely(thread_is_stopped(thread) &&
1569 list_empty(&sai->sai_entries_stated))) {
1570 /* to release resource */
1571 ll_stop_statahead(dir, lli->lli_opendir_key);
1575 if ((*dentryp)->d_name.name[0] == '.') {
1576 if (sai->sai_ls_all ||
1577 sai->sai_miss_hidden >= sai->sai_skip_hidden) {
1579 * Hidden dentry is the first one, or statahead
1580 * thread does not skip so many hidden dentries
1581 * before "sai_ls_all" enabled as below.
1584 if (!sai->sai_ls_all)
1586 * It maybe because hidden dentry is not
1587 * the first one, "sai_ls_all" was not
1588 * set, then "ls -al" missed. Enable
1589 * "sai_ls_all" for such case.
1591 sai->sai_ls_all = 1;
1594 * Such "getattr" has been skipped before
1595 * "sai_ls_all" enabled as above.
1597 sai->sai_miss_hidden++;
1602 entry = ll_sa_entry_get_byname(sai, &(*dentryp)->d_name);
1603 if (entry == NULL || only_unplug) {
1604 ll_sai_unplug(sai, entry);
1605 RETURN(entry ? 1 : -EAGAIN);
1608 /* if statahead is busy in readdir, help it do post-work */
1609 while (!ll_sa_entry_stated(entry) &&
1610 sai->sai_in_readpage && !sa_received_empty(sai))
1611 ll_post_statahead(sai);
1613 if (!ll_sa_entry_stated(entry)) {
1614 sai->sai_index_wait = entry->se_index;
1615 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(30), NULL,
1616 LWI_ON_SIGNAL_NOOP, NULL);
1617 rc = l_wait_event(sai->sai_waitq,
1618 ll_sa_entry_stated(entry) ||
1619 thread_is_stopped(thread),
1622 ll_sai_unplug(sai, entry);
1627 if (entry->se_stat == SA_ENTRY_SUCC &&
1628 entry->se_inode != NULL) {
1629 struct inode *inode = entry->se_inode;
1630 struct lookup_intent it = { .it_op = IT_GETATTR,
1631 .d.lustre.it_lock_handle =
1635 rc = md_revalidate_lock(ll_i2mdexp(dir), &it,
1636 ll_inode2fid(inode), &bits);
1638 if ((*dentryp)->d_inode == NULL) {
1639 struct dentry *alias;
1641 alias = ll_splice_alias(inode,
1643 if (IS_ERR(alias)) {
1644 ll_sai_unplug(sai, entry);
1645 RETURN(PTR_ERR(alias));
1648 } else if ((*dentryp)->d_inode != inode) {
1649 /* revalidate, but inode is recreated */
1651 "%s: stale dentry %.*s inode "
1652 DFID", statahead inode "DFID
1654 ll_get_fsname((*dentryp)->d_inode->i_sb, NULL, 0),
1655 (*dentryp)->d_name.len,
1656 (*dentryp)->d_name.name,
1657 PFID(ll_inode2fid((*dentryp)->d_inode)),
1658 PFID(ll_inode2fid(inode)));
1659 ll_intent_release(&it);
1660 ll_sai_unplug(sai, entry);
1665 entry->se_inode = NULL;
1667 if ((bits & MDS_INODELOCK_LOOKUP) &&
1668 d_lustre_invalid(*dentryp))
1669 d_lustre_revalidate(*dentryp);
1670 ll_intent_release(&it);
1674 ll_sai_unplug(sai, entry);
1678 /* I am the "lli_opendir_pid" owner, only me can set "lli_sai". */
1679 rc = is_first_dirent(dir, *dentryp);
1680 if (rc == LS_NONE_FIRST_DE)
1681 /* It is not "ls -{a}l" operation, no need statahead for it. */
1682 GOTO(out, rc = -EAGAIN);
1684 sai = ll_sai_alloc();
1686 GOTO(out, rc = -ENOMEM);
1688 sai->sai_ls_all = (rc == LS_FIRST_DOT_DE);
1689 sai->sai_inode = igrab(dir);
1690 if (unlikely(sai->sai_inode == NULL)) {
1691 CWARN("Do not start stat ahead on dying inode "DFID"\n",
1692 PFID(&lli->lli_fid));
1693 GOTO(out, rc = -ESTALE);
1696 /* get parent reference count here, and put it in ll_statahead_thread */
1697 parent = dget((*dentryp)->d_parent);
1698 if (unlikely(sai->sai_inode != parent->d_inode)) {
1699 struct ll_inode_info *nlli = ll_i2info(parent->d_inode);
1701 CWARN("Race condition, someone changed %.*s just now: "
1702 "old parent "DFID", new parent "DFID"\n",
1703 (*dentryp)->d_name.len, (*dentryp)->d_name.name,
1704 PFID(&lli->lli_fid), PFID(&nlli->lli_fid));
1706 iput(sai->sai_inode);
1707 GOTO(out, rc = -EAGAIN);
1710 CDEBUG(D_READA, "start statahead thread: sai %p, parent %.*s\n",
1711 sai, parent->d_name.len, parent->d_name.name);
1713 /* The sai buffer already has one reference taken at allocation time,
1714 * but as soon as we expose the sai by attaching it to the lli that
1715 * default reference can be dropped by another thread calling
1716 * ll_stop_statahead. We need to take a local reference to protect
1717 * the sai buffer while we intend to access it. */
1721 plli = ll_i2info(parent->d_inode);
1722 task = kthread_run(ll_statahead_thread, parent, "ll_sa_%u",
1723 plli->lli_opendir_pid);
1724 thread = &sai->sai_thread;
1727 CERROR("cannot start ll_sa thread: rc = %d\n", rc);
1729 lli->lli_opendir_key = NULL;
1730 thread_set_flags(thread, SVC_STOPPED);
1731 thread_set_flags(&sai->sai_agl_thread, SVC_STOPPED);
1732 /* Drop both our own local reference and the default
1733 * reference from allocation time. */
1736 LASSERT(lli->lli_sai == NULL);
1740 l_wait_event(thread->t_ctl_waitq,
1741 thread_is_running(thread) || thread_is_stopped(thread),
1746 * We don't stat-ahead for the first dirent since we are already in
1754 spin_lock(&lli->lli_sa_lock);
1755 lli->lli_opendir_key = NULL;
1756 lli->lli_opendir_pid = 0;
1757 spin_unlock(&lli->lli_sa_lock);